<p>This study investigates reservoir computing for predicting chaotic systems from highly incomplete data. To address missing data, a sparse reconstruction method is employed to impute missing values and ensure the completeness of the information flow. Furthermore, a sample-weighted reservoir computing model is proposed to regulate the contributions of error-containing imputed data and accurate observed data during the computation of the readout weight matrix. Experimental verification on benchmark systems, including the Sprott-A, Dadras-Momeni and Kuramoto–Sivashinsky systems, shows that even when the proportion of missing data reaches or exceeds 85%, the trained model is able to capture the underlying chaotic dynamics and make short-term accurate prediction as well as long-term climate prediction of the system's evolution. For a parameterized second-order time-delay system, a model constructed with sparse observed data from only a few parameter values successfully predicts multiple bifurcation diagrams. Our results demonstrate that effective prediction of chaotic dynamics can be achieved using only a small, randomly distributed subset of data points from the full time series.</p>